| #!/usr/bin/env python3 | |
| """Layer-weighted delta-scale merge, grounded in the measured weight geometry: | |
| Coder DESCENDS from V2 (late-layer proj->1.0), its delta is ~1.8x v2's magnitude, and in EARLY | |
| layers the coding delta is ANTI-aligned with v2's reasoning (cos<0) => it damaged reasoning/control. | |
| So: W_new = V2 + alpha(L) * (Coder - V2), alpha ramps low (early, protect reasoning) -> high (late, | |
| inject coding). embed/lm_head/norm/mtp/vision are frozen across all 3 -> copied from V2 verbatim. | |
| """ | |
| import os, re, json, gc, sys, argparse | |
| import torch | |
| from safetensors import safe_open | |
| from safetensors.torch import save_file | |
| SSD="/media/kylehessling/Extreme SSD" | |
| V2_PATH=f"{SSD}/Qwopus3.6-27B-v2" # base (reasoning) — also donates embed/lm_head/norm/mtp/vision | |
| CO_PATH=f"{SSD}/Qwopus3.6-27B-Coder" # coder — delta source | |
| SHARD=5_000_000_000 | |
| NLAYERS=64 | |
| COPY_V2=re.compile(r"(embed_tokens|lm_head|^mtp\.|\.mtp\.|nextn|\.norm\.|vision|visual|image|video|patch_embed|merger)") | |
| LAYER_RE=re.compile(r"\.layers\.(\d+)\.") | |
| def load_index(p): | |
| return {k:os.path.join(p,v) for k,v in json.load(open(f"{p}/model.safetensors.index.json"))["weight_map"].items()} | |
| class ShardWriter: | |
| def __init__(s,o): s.o,s.buf,s.cur,s.i,s.tot,s.wm=o,{},0,0,0,{}; os.makedirs(o,exist_ok=True) | |
| def _flush(s): | |
| if not s.buf: return | |
| s.i+=1; nm=f"model-{s.i:05d}.safetensors"; save_file(s.buf,os.path.join(s.o,nm),metadata={"format":"pt"}) | |
| for k in s.buf: s.wm[k]=nm | |
| s.buf,s.cur={},0; gc.collect() | |
| def add(s,n,t): s.buf[n]=t.contiguous(); s.cur+=t.numel()*t.element_size(); s.tot+=1; (s._flush() if s.cur>=SHARD else None) | |
| def finalize(s): | |
| s._flush(); n=s.i | |
| for j in range(1,n+1): | |
| old=f"model-{j:05d}.safetensors"; new=f"model-{j:05d}-of-{n:05d}.safetensors" | |
| os.rename(os.path.join(s.o,old),os.path.join(s.o,new)) | |
| for k,v in list(s.wm.items()): | |
| if v==old: s.wm[k]=new | |
| json.dump({"metadata":{"total_size":0},"weight_map":s.wm},open(f"{s.o}/model.safetensors.index.json","w"),indent=2) | |
| def alpha(L, a_early, a_late): | |
| return a_early + (a_late-a_early)*(L/(NLAYERS-1)) # smooth linear ramp by depth | |
| def main(): | |
| ap=argparse.ArgumentParser() | |
| ap.add_argument("--out",required=True) | |
| ap.add_argument("--a-early",type=float,default=0.12) | |
| ap.add_argument("--a-late",type=float,default=0.48) | |
| ap.add_argument("--uniform",type=float,default=None,help="if set, constant alpha (control candidate)") | |
| a=ap.parse_args() | |
| iv,ic=load_index(V2_PATH),load_index(CO_PATH) | |
| assert set(iv)==set(ic),"key mismatch" | |
| H={} | |
| def get(i,n): | |
| p=i[n] | |
| if p not in H: H[p]=safe_open(p,"pt") | |
| return H[p].get_tensor(n) | |
| w=ShardWriter(a.out); copied=merged=0 | |
| for i,n in enumerate(sorted(iv)): | |
| if COPY_V2.search(n): | |
| w.add(n,get(iv,n)); copied+=1 | |
| else: | |
| m=LAYER_RE.search(n); L=int(m.group(1)) if m else NLAYERS//2 | |
| al = a.uniform if a.uniform is not None else alpha(L,a.a_early,a.a_late) | |
| v2=get(iv,n).float(); co=get(ic,n).float() | |
| out=(v2 + al*(co-v2)).to(torch.bfloat16) | |
| w.add(n,out); merged+=1 | |
| if i%150==0: | |
| print(f" [{i}/{len(iv)}] copied={copied} merged={merged}",flush=True); H.clear(); gc.collect() | |
| w.finalize() | |
| sched = f"uniform {a.uniform}" if a.uniform is not None else f"ramp {a.a_early}->{a.a_late}" | |
| print(f"[merge] DONE ({sched}) copied={copied} merged={merged} -> {a.out}",flush=True) | |
| import shutil | |
| for fn in os.listdir(V2_PATH): | |
| if fn.endswith((".json",".model",".txt")) and "safetensors" not in fn: shutil.copy2(os.path.join(V2_PATH,fn),os.path.join(a.out,fn)) | |
| print("[merge] copied config+tokenizer from V2",flush=True) | |
| if __name__=="__main__": main() | |
Xet Storage Details
- Size:
- 3.85 kB
- Xet hash:
- 4945ebfaa38c86947215d1dd274da6c6d63ab1802b02a91a49d39592c4e7c818
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.