| """Pre-merge diagnostic ONLY, for all forks: is the fork still in the base model's frame? |
| Writes the diag records + the fitted g so chatvec_run.py reuses them.""" |
| import os, sys, json, time, gc, traceback |
| os.environ["CUDA_VISIBLE_DEVICES"] = sys.argv[2] |
| import numpy as np, torch |
| import ma_common as C, gmap |
| from mergeschool.core import metrics as MT |
|
|
| FORKS = json.load(open(sys.argv[1])) |
| LEDGER = os.environ.get("MA_LEDGER", "/root/merge-accuracy/results/chatvec.jsonl") |
| BASE = "meta-llama/Llama-3.1-8B" |
| done = C.jload(LEDGER) |
|
|
| mb = C.load_model(BASE, dev="cpu", dtype=torch.float32) |
| sd_base = C.sd_np(mb); cfg = mb.config |
| HID, NH = cfg.hidden_size, cfg.num_attention_heads |
| NKV = getattr(cfg, "num_key_value_heads", NH) |
| del mb; gc.collect() |
| tok_base = C.load_tok(BASE) |
| sents = C.flores_lines("eng_Latn", 256) |
| m = C.load_model(BASE, dev="cuda", dtype=torch.bfloat16) |
| acts_base = C.capture_acts_sent(m, tok_base, sents, "cuda") |
| del m; gc.collect(); torch.cuda.empty_cache() |
| print("base ready", flush=True) |
|
|
| for F in FORKS: |
| name, repo, lang = F["name"], F["repo"], F["lang"] |
| if f"{name}|diag" in done: continue |
| try: |
| t0 = time.time() |
| tok_f = C.load_tok(repo) |
| mfg = C.load_model(repo, dev="cuda", dtype=torch.bfloat16) |
| acts_f = C.capture_acts_sent(mfg, tok_f, sents, "cuda") |
| del mfg; gc.collect(); torch.cuda.empty_cache() |
| mf = C.load_model(repo, dev="cpu", dtype=torch.float32) |
| sd_fork = C.sd_np(mf); del mf; gc.collect() |
| KEYS = C.shared_keys(sd_base, sd_fork) |
| g, info = gmap.fit_g(sd_fork, sd_base, HID, NH, acts_f, acts_base, "permutation", verbose=False, n_kv_heads=NKV) |
| a = np.concatenate([sd_base[k].ravel() for k in KEYS]) |
| b = np.concatenate([sd_fork[k].ravel() for k in KEYS]) |
| info["weight_cosine_vs_base"] = float(a @ b / (np.linalg.norm(a) * np.linalg.norm(b))) |
| info["rel_drift"] = float(np.linalg.norm(a - b) / np.linalg.norm(a)) |
| del a, b; gc.collect() |
| L = sorted(set(acts_f) & set(acts_base)) |
| info["cka_mean"] = float(np.mean([MT.cka(acts_base[l], acts_f[l]) for l in L])) |
| info["cka_last"] = float(MT.cka(acts_base[L[-1]], acts_f[L[-1]])) |
| info["coord_share"] = info["coord_share_bn"] |
| info["PREDICTION_align_helps"] = bool(info["coord_share"] >= 0.01) |
| info["fit_seconds"] = time.time() - t0 |
| info = {k: (v.tolist() if isinstance(v, np.ndarray) else v) for k, v in info.items()} |
| np.save(f"/root/merge-accuracy/results/g_{name}.npy", np.array([g], dtype=object), allow_pickle=True) |
| rec = {"key": f"{name}|diag", "kind": "diag", "fork": name, "arm": "diag", "lang": lang, |
| "t": time.time(), "diag": info} |
| C.jappend(LEDGER, rec) |
| print(f"DIAG {name}: coord_share={info['coord_share']:.5f} identity={info['is_identity']} " |
| f"bnd_raw={info['bnd_raw']:.4f} bnd_final={info['bnd_final']:.4f} " |
| f"wcos={info['weight_cosine_vs_base']:.4f} drift={info['rel_drift']:.4f} " |
| f"cka={info['cka_mean']:.3f} residual={info['residual']} hidden={info['hidden']} " |
| f"heads={info['heads']} rejected={info['rejected']} {info['fit_seconds']:.0f}s", flush=True) |
| del sd_fork, acts_f, g; gc.collect() |
| except Exception: |
| print(f"!! DIAG FAIL {name}\n{traceback.format_exc()[-1500:]}", flush=True) |
| print("DIAG_DONE", flush=True) |
|
|
