layerlab-store / code /ck_discover.py
cybertronak's picture
before CircuitKIT HF job (patch_groups.json + first_tokens.json must be in the store)
45bf534 verified
Raw History Blame Contribute Delete
2.04 kB
import sys, json, re, math, torch
import circuitkit as ck
algo, n, dtype, out, pairs, refuse, comply = sys.argv[1], int(sys.argv[2]), sys.argv[3], sys.argv[4], sys.argv[5], sys.argv[6], sys.argv[7]
model = ck.load_model("CohereLabs/tiny-aya-global", dtype=dtype)
for name, p in model.named_parameters(): p.requires_grad_(name.endswith((".w", ".b")) and "ln" in name)
torch.cuda.empty_cache(); print("model loaded | GB allocated:", round(torch.cuda.memory_allocated() / 1e9, 2), "| trainable params (M):", round(sum(p.numel() for p in model.parameters() if p.requires_grad) / 1e6, 3), flush=True)
dp = {"language": "en", "pairs_path": pairs, "refuse": refuse, "comply": comply, "num_examples": n}
try:
circuit = ck.discover(model, "multilingual_refusal", algorithm=algo, n_examples=n, batch_size=1, sparsity=0.3, scope="both", output_path=out + "/circuit.pt", chat_template_mode="off", data_params=dp)
except Exception:
print("peak GB at failure:", round(torch.cuda.max_memory_allocated() / 1e9, 2), flush=True); raise
vals = [float(v) for v in circuit.scores.values()]
if not vals or any(not math.isfinite(v) for v in vals) or max(abs(v) for v in vals) == 0:
print("DEGENERATE SCORES: n =", len(vals), "| non-finite =", sum(not math.isfinite(v) for v in vals)); sys.exit(3)
layers = {}
for name, s in circuit.scores.items():
m = re.match(r"[Aa](\d+)\.[hH]?(\d+)$", name); kind = "attn"
if not m: m = re.match(r"MLP\s*(\d+)$", name, re.I); kind = "mlp"
if not m: continue
d = layers.setdefault(int(m.group(1)), {"attn": 0.0, "mlp": 0.0}); d[kind] += abs(float(s))
if not layers:
print("NO LAYERS PARSED; example score keys:", list(circuit.scores)[:8]); sys.exit(2)
json.dump({"algo": algo, "dtype": dtype, "n_examples": n, "layers": layers, "n_nodes": len(circuit.scores), "peak_gb": torch.cuda.max_memory_allocated() / 1e9}, open(out + "/ck_layer_scores.json", "w"))
print("layers scored:", len(layers), "| nodes:", len(circuit.scores), "| peak GB:", round(torch.cuda.max_memory_allocated() / 1e9, 2))