| import io, json, os, zipfile, resource |
| import numpy as np, torch |
| from safetensors import safe_open |
| from torch.distributed.checkpoint import FileSystemReader |
| CKPT="/projects/u6uv/ziyan/models/megatron/Qwen3.8-27B_torch_dist/release" |
| HF="/projects/u6uv/ziyan/models/Qwen/Qwen3.8-27B" |
| idx=json.load(open(os.path.join(HF,"model.safetensors.index.json")))["weight_map"] |
| def rssG(): return resource.getrusage(resource.RUSAGE_SELF).ru_maxrss/1e6 |
|
|
| class Slice(io.RawIOBase): |
| def __init__(s,p,o,l): s.f=open(p,'rb'); s.o=o; s.l=l; s.p=0 |
| def readable(s): return True |
| def seekable(s): return True |
| def seek(s,off,wh=0): |
| s.p = off if wh==0 else (s.p+off if wh==1 else s.l+off); return s.p |
| def tell(s): return s.p |
| def read(s,n=-1): |
| if n is None or n<0: n=s.l-s.p |
| n=min(n,s.l-s.p); s.f.seek(s.o+s.p); b=s.f.read(n); s.p+=len(b); return b |
| def readinto(s,b): |
| d=s.read(len(b)); b[:len(d)]=d; return len(d) |
|
|
| md=FileSystemReader(CKPT).read_metadata() |
| targets={"embedding.word_embeddings.weight":"model.language_model.embed_tokens.weight", |
| "output_layer.weight":"lm_head.weight"} |
| si={} |
| for k,v in md.storage_data.items(): |
| if getattr(k,"fqn",None) in targets: si[k.fqn]=v |
|
|
| for mn,hn in targets.items(): |
| info=si[mn] |
| zf=zipfile.ZipFile(io.BufferedReader(Slice(os.path.join(CKPT,info.relative_path),info.offset,info.length))) |
| names=[n for n in zf.namelist() if "/data/" in n or n.endswith("data/0")] |
| dataname=[n for n in zf.namelist() if n.split("/")[-1].isdigit()] |
| print(mn, "zip entries:", [(n, zf.getinfo(n).file_size) for n in zf.namelist()][:6], flush=True) |
| dn=max(dataname, key=lambda n: zf.getinfo(n).file_size) |
| total=zf.getinfo(dn).file_size |
| rows, cols = 248320, 5120 |
| assert total==rows*cols*2, (total, rows*cols*2) |
| stream=zf.open(dn) |
| with safe_open(os.path.join(HF, idx[hn]), framework="pt", device="cpu") as g: |
| sl=g.get_slice(hn) |
| nb=0; done=0; CH=4096 |
| while done<rows: |
| n=min(CH, rows-done) |
| raw=stream.read(n*cols*2) |
| a=torch.from_numpy(np.frombuffer(raw, dtype=np.int16).copy()).reshape(n,cols) |
| ref=sl[done:done+n].contiguous().view(torch.int16) |
| nb+=int((a!=ref).sum()) |
| done+=n; del raw,a,ref |
| print(f" {mn} -> {hn}: in_hf={hn in idx} rows={rows} nbitdiff={nb} maxRSS_GB={rssG():.2f}", flush=True) |
| stream.close(); zf.close() |
|
|
