File size: 5,001 Bytes
fb9e149 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 | """Broad ecosystem screen: how many RELEASED Llama-3.1-8B derivatives have actually left the base
model's coordinate frame? Streams one model at a time (download -> screen -> delete) so the disk
footprint stays at one checkpoint."""
import os, sys, json, time, gc, shutil, traceback
sys.path.insert(0, "/root/merge-accuracy")
import numpy as np, torch
from huggingface_hub import snapshot_download, hf_hub_download
import ma_common as C, gmap
from cheap_screen import screen
from mergeschool.core import alignment as AL
OUT = "/root/merge-accuracy/results/ecosystem_screen.json"
CACHE = "/root/hf_cache_mergeacc"
BASE = "meta-llama/Llama-3.1-8B"
CAND = [
("NousResearch/Hermes-3-Llama-3.1-8B", "instruct post-training", "Nous Research"),
("allenai/Llama-3.1-Tulu-3-8B-SFT", "instruct SFT", "AI2"),
("dphn/Dolphin3.0-Llama3.1-8B", "instruct post-training", "Dolphin"),
("meta-llama/Llama-Guard-3-8B", "safety classifier", "Meta"),
("fdtn-ai/Foundation-Sec-8B", "domain CPT (security)", "Foundation AI"),
("OpenSciLM/Llama-3.1_OpenScholar-8B", "domain CPT (science)", "OpenSciLM"),
("nvidia/OpenMath2-Llama3.1-8B", "domain SFT (maths)", "NVIDIA"),
("NCSOFT/Llama-VARCO-8B-Instruct", "language CPT (Korean)", "NCSOFT"),
("McGill-NLP/AfriqueLlama-8B", "language CPT (African)", "McGill NLP"),
("Yiddish-NLP/MameLoshnLM", "language CPT (Yiddish)", "Yiddish-NLP"),
("deepcogito/cogito-v1-preview-llama-8B", "instruct post-training", "Deep Cogito"),
("tokyotech-llm/Llama-3.1-Swallow-8B-v0.2", "language CPT (Japanese)", "TokyoTech"),
("aisingapore/Llama-SEA-LION-v3-8B", "language CPT (SEA)", "AI Singapore"),
("microsoft/UserLM-8b", "role post-training", "Microsoft"),
]
res = json.load(open(OUT)) if os.path.exists(OUT) else {}
mb = C.load_model(BASE, dev="cpu", dtype=torch.float32)
sd_base = C.sd_np(mb); cfg = mb.config
HID, NH, NKV, VOC = cfg.hidden_size, cfg.num_attention_heads, cfg.num_key_value_heads, cfg.vocab_size
del mb; gc.collect()
print(f"base ready HID={HID} NH={NH} NKV={NKV}", flush=True)
for repo, kind, group in CAND:
if repo in res: continue
d = None
try:
cfp = hf_hub_download(repo, "config.json", cache_dir=CACHE)
c = json.load(open(cfp))
if (c.get("hidden_size") != HID or c.get("num_attention_heads") != NH
or c.get("num_key_value_heads") != NKV or c.get("vocab_size") != VOC
or c.get("num_hidden_layers") != cfg.num_hidden_layers):
res[repo] = {"kind": kind, "group": group, "status": "shape mismatch",
"config": {k: c.get(k) for k in ("hidden_size", "num_attention_heads",
"num_key_value_heads", "vocab_size",
"num_hidden_layers")}}
print(f"SKIP {repo}: shape mismatch", flush=True)
json.dump(res, open(OUT, "w"), indent=1); continue
t0 = time.time()
d = snapshot_download(repo, allow_patterns=["*.safetensors", "*.json", "tokenizer*"],
cache_dir=CACHE, max_workers=8)
dl = time.time() - t0
m = C.load_model(repo, dev="cpu", dtype=torch.float32)
sd = C.sd_np(m); del m; gc.collect()
t = time.time(); f_id, pres = screen(sd, sd_base, HID); dt = time.time() - t
keys = C.shared_keys(sd_base, sd)
a = np.concatenate([sd_base[k].ravel() for k in keys])
b = np.concatenate([sd[k].ravel() for k in keys])
wc = float(a @ b / (np.linalg.norm(a) * np.linalg.norm(b)))
rd = float(np.linalg.norm(a - b) / np.linalg.norm(a))
del a, b, sd; gc.collect()
res[repo] = {"kind": kind, "group": group, "status": "screened",
"identity_fraction_worst_layer": f_id, "screen_seconds": dt,
"download_seconds": dl, "layers_screened": len(pres),
"weight_cosine_vs_base": wc, "rel_drift": rd,
"frame_has_drifted": bool(f_id < 0.95)}
print(f"{repo:48s} id_frac={f_id:.4f} wcos={wc:.4f} drift={rd:.4f} "
f"-> {'DRIFTED' if f_id < 0.95 else 'same frame'} ({dt:.0f}s)", flush=True)
json.dump(res, open(OUT, "w"), indent=1)
except Exception:
res[repo] = {"kind": kind, "group": group, "status": "error",
"error": traceback.format_exc()[-400:]}
print(f"ERR {repo}: {traceback.format_exc()[-250:]}", flush=True)
json.dump(res, open(OUT, "w"), indent=1)
finally:
# stream and delete: keep at most one candidate checkpoint on disk
keep = ("Llama-3.1-8B", "Llama-3.1-8B-Instruct", "Swallow-8B-v0.1",
"typhoon2", "sea-lionv3-base")
if not any(k in repo for k in keep):
p = f"{CACHE}/models--" + repo.replace("/", "--")
if os.path.isdir(p):
shutil.rmtree(p, ignore_errors=True)
gc.collect()
print("SWEEP_DONE", flush=True)
|