"""Reverse-flossing fine-tune of a trained HRM checkpoint. L_total = L_HRM_ACT (existing supervised loss) + alpha * L_RF L_RF = mean over batch of max(0, mean_lambda - lambda_star) ** 2 Computes finite-time top-k Lyapunov spectrum during the forward (with grad enabled so that L_RF is differentiable wrt model params), then averages and applies a hinge. """ from __future__ import annotations import sys, os, yaml, json, math, time, argparse from pathlib import Path import numpy as np import torch import torch.nn.functional as F HRM_DIR = Path("/home/yurenh2/rrm/hrm") sys.path.insert(0, str(HRM_DIR)) from models.hrm.hrm_act_v1 import ( HierarchicalReasoningModel_ACTV1, HierarchicalReasoningModel_ACTV1InnerCarry, ) from models.losses import ACTLossHead from adam_atan2 import AdamATan2 # ----------------- helpers ----------------- def load_model_for_finetune(ckpt_root: Path, ckpt_name: str, device: str): cfg = yaml.safe_load((ckpt_root / "all_config.yaml").read_text()) arch_cfg = dict(cfg["arch"]) train_meta = json.loads((Path(cfg["data_path"]) / "train" / "dataset.json").read_text()) arch_cfg.update( batch_size=cfg["global_batch_size"], seq_len=train_meta["seq_len"], vocab_size=train_meta["vocab_size"], num_puzzle_identifiers=train_meta["num_puzzle_identifiers"], causal=False, ) base = HierarchicalReasoningModel_ACTV1(arch_cfg) head = ACTLossHead(base, loss_type=arch_cfg["loss"]["loss_type"]) sd = torch.load(ckpt_root / ckpt_name, map_location="cpu", weights_only=True) stripped = {} for k, v in sd.items(): nk = k for prefix in ("_orig_mod.",): if nk.startswith(prefix): nk = nk[len(prefix):] stripped[nk] = v missing, unexpected = head.load_state_dict(stripped, strict=False) print(f"[load] missing={len(missing)} unexpected={len(unexpected)}") if missing[:3]: print(f" sample missing: {missing[:3]}") if unexpected[:3]: print(f" sample unexpected: {unexpected[:3]}") head.to(device) return head, base, cfg, train_meta def jvp_apply(f, x, V): """D_f(x) @ V where V is (B, state_dim, k). Uses create_graph=True so RF loss is differentiable wrt model params.""" B, state_dim, k = V.shape out = [] fx_last = None for i in range(k): v_i = V[..., i].view_as(x) fx, Dv = torch.autograd.functional.jvp(f, x, v=v_i, create_graph=True, strict=False) out.append(Dv.reshape(B, state_dim).to(torch.float32)) fx_last = fx return fx_last, torch.stack(out, dim=-1) def compute_lyap_mean(model, inner, batch, k_lyap, device, seed, lyap_act_steps=4): """Run forward with grad and accumulate top-k Lyapunov estimate per sample. Returns mean_lambda per sample (B,) — DIFFERENTIABLE wrt params. `lyap_act_steps`: how many ACT steps to unroll for the Lyapunov estimate. Fewer steps → lower memory but noisier estimate. T = lyap_act_steps * (H*L + H). """ cfg = inner.config B = batch["inputs"].shape[0] seq_full = inner.config.seq_len + inner.puzzle_emb_len hidden = cfg.hidden_size state_dim = seq_full * hidden # Init z_H = inner.H_init.unsqueeze(0).expand(B, seq_full, hidden).clone().to(inner.forward_dtype) z_L = inner.L_init.unsqueeze(0).expand(B, seq_full, hidden).clone().to(inner.forward_dtype) seq_info = dict(cos_sin=inner.rotary_emb() if hasattr(inner, "rotary_emb") else None) input_embeddings = inner._input_embeddings(batch["inputs"], batch["puzzle_identifiers"]) g = torch.Generator(device=device).manual_seed(seed) Q0 = torch.randn(B, state_dim, k_lyap, device=device, dtype=torch.float32, generator=g) Q, _ = torch.linalg.qr(Q0) log_R_sum = torch.zeros(B, k_lyap, device=device, dtype=torch.float32) n_steps = 0 n_act = min(lyap_act_steps, cfg.halt_max_steps) for _act in range(n_act): for _h in range(cfg.H_cycles): for _l in range(cfg.L_cycles): f = lambda x: inner.L_level(x, z_H + input_embeddings, **seq_info) z_L_new, DQ = jvp_apply(f, z_L, Q) Q = DQ; z_L = z_L_new Q, R = torch.linalg.qr(Q) log_R_sum = log_R_sum + R.diagonal(dim1=-2, dim2=-1).abs().clamp_min(1e-30).log() n_steps += 1 f = lambda x: inner.H_level(x, z_L, **seq_info) z_H_new, DQ = jvp_apply(f, z_H, Q) Q = DQ; z_H = z_H_new Q, R = torch.linalg.qr(Q) log_R_sum = log_R_sum + R.diagonal(dim1=-2, dim2=-1).abs().clamp_min(1e-30).log() n_steps += 1 lyap_spec = log_R_sum / max(n_steps, 1) # (B, k) return lyap_spec.mean(dim=-1) # (B,) mean over top-k # ----------------- data ----------------- def load_train_batches(data_path: Path, batch_size: int, n_iters: int, seed: int = 0): """Tiny iterator over the augmented train set.""" rng = np.random.default_rng(seed) inputs = np.load(data_path / "train" / "all__inputs.npy") labels = np.load(data_path / "train" / "all__labels.npy") pid = np.load(data_path / "train" / "all__puzzle_identifiers.npy") N = len(inputs) for _ in range(n_iters): idx = rng.choice(N, size=batch_size, replace=False) yield { "inputs": torch.from_numpy(inputs[idx].astype(np.int32)), "labels": torch.from_numpy(labels[idx].astype(np.int32)), "puzzle_identifiers": torch.from_numpy(pid[idx].astype(np.int32)), } def evaluate(head, base, data_path: Path, n_samples: int, batch_size: int, device: str, seed: int = 42): """Quick exact-accuracy eval on test set.""" rng = np.random.default_rng(seed) inputs = np.load(data_path / "test" / "all__inputs.npy") labels = np.load(data_path / "test" / "all__labels.npy") pid = np.load(data_path / "test" / "all__puzzle_identifiers.npy") idx_all = rng.choice(len(inputs), size=n_samples, replace=False) head.eval() correct = 0; token_correct = 0; token_total = 0 for s in range(0, n_samples, batch_size): e = min(s + batch_size, n_samples) idx = idx_all[s:e] batch = { "inputs": torch.from_numpy(inputs[idx].astype(np.int32)).to(device), "labels": torch.from_numpy(labels[idx].astype(np.int32)).to(device), "puzzle_identifiers": torch.from_numpy(pid[idx].astype(np.int32)).to(device), } with torch.no_grad(): with torch.device(device): carry = base.initial_carry(batch) for _ in range(base.config.halt_max_steps): carry, outputs = base(carry=carry, batch=batch) preds = outputs["logits"].argmax(dim=-1) mask = batch["labels"] > 0 exact = ((preds == batch["labels"]) | ~mask).all(dim=-1).float() correct += exact.sum().item() token_correct += ((preds == batch["labels"]) & mask).sum().item() token_total += mask.sum().item() return correct / n_samples, token_correct / max(token_total, 1) # ----------------- main ----------------- def main(): ap = argparse.ArgumentParser() ap.add_argument("--ckpt-root", required=True) ap.add_argument("--ckpt-name", default="step_26040") ap.add_argument("--n-steps", type=int, default=200) ap.add_argument("--batch-size", type=int, default=16, help="kept small because RF JVPs use a lot of memory") ap.add_argument("--lr", type=float, default=2e-5) ap.add_argument("--alpha-rf", type=float, default=1.0) ap.add_argument("--lambda-star", type=float, default=-0.85) ap.add_argument("--k-lyap", type=int, default=2) ap.add_argument("--lyap-act-steps", type=int, default=4, help="ACT steps to unroll for Lyapunov measurement") ap.add_argument("--rf-mode", choices=["fixed","horizon"], default="fixed", help="fixed: use --lambda-star as target. horizon: λ*=(1/T)log(eps/loss)") ap.add_argument("--rf-eps", type=float, default=1e-6, help="task tolerance for horizon mode") ap.add_argument("--seed", type=int, default=0) ap.add_argument("--eval-every", type=int, default=50) ap.add_argument("--eval-n", type=int, default=512) ap.add_argument("--eval-batch-size", type=int, default=64) ap.add_argument("--out", default="rf_finetune_log.json") args = ap.parse_args() device = "cuda" head, base, cfg, train_meta = load_model_for_finetune(Path(args.ckpt_root), args.ckpt_name, device) # Optimizer matching HRM training optim = AdamATan2(head.parameters(), lr=args.lr, betas=(0.9, 0.95), weight_decay=cfg["weight_decay"]) # Initial eval print(f"\n=== Initial eval (no fine-tune) ===") acc0, tacc0 = evaluate(head, base, Path(cfg["data_path"]), args.eval_n, args.eval_batch_size, device) print(f" initial exact_acc = {acc0:.4f} token_acc = {tacc0:.4f}") log = {"args": vars(args), "initial_acc": acc0, "initial_tok_acc": tacc0, "steps": []} t0 = time.time() train_iter = load_train_batches(Path(cfg["data_path"]), args.batch_size, args.n_steps, seed=args.seed) head.train() # Keep sparse puzzle embedding in eval mode (its local_weights buffer is sized for # the original training batch_size of 768; here we use a smaller batch). The # puzzle_emb table is still in the model, just not updated during fine-tune. base.inner.puzzle_emb.eval() for p in base.inner.puzzle_emb.parameters(): p.requires_grad_(False) for step, batch in enumerate(train_iter): batch = {k: v.to(device) for k, v in batch.items()} # Make sure model is in train mode but puzzle_emb stays in eval mode (its # local_weights buffer is sized for the original 768 batch_size; we use a smaller batch). head.train() base.inner.puzzle_emb.eval() # ---- Supervised ACT loss accumulated over all halt_max_steps ACT steps ---- # We use a fresh carry per fine-tune step; run ACT loop fully and sum losses. with torch.device(device): carry = base.initial_carry(batch) sup_loss_sum = 0.0 n_loss = 0 for _ in range(base.config.halt_max_steps): carry, l, metrics, _, all_finish = head(return_keys=[], carry=carry, batch=batch) sup_loss_sum = sup_loss_sum + l n_loss += 1 if all_finish: break # Average across ACT steps, then normalize by batch_size sup_loss = sup_loss_sum / max(n_loss, 1) / args.batch_size # ---- Reverse-flossing penalty ---- # New forward pass dedicated to computing finite-time mean λ (differentiable). mean_lyap = compute_lyap_mean(head, base.inner, batch, args.k_lyap, device, seed=args.seed + step, lyap_act_steps=args.lyap_act_steps) # λ̄ is in nats per inner-cycle. if args.rf_mode == "fixed": lam_star = torch.tensor(args.lambda_star, device=device, dtype=torch.float32) else: # finite-horizon: λ* = (1/T) log(eps / r); r = current per-sample sup loss proxy (use total loss) T = base.config.halt_max_steps * (base.config.H_cycles * base.config.L_cycles + base.config.H_cycles) r = sup_loss.detach().clamp_min(1e-9) lam_star = (1.0 / T) * torch.log(torch.tensor(args.rf_eps, device=device) / r) # Cap at small negative so we don't ask for impossible contraction lam_star = lam_star.clamp(min=-2.0, max=0.0) excess = (mean_lyap - lam_star).clamp_min(0.0) # 0 if lyap < star (already contractive enough) rf_loss = (excess ** 2).mean() total_loss = sup_loss + args.alpha_rf * rf_loss optim.zero_grad(set_to_none=True) total_loss.backward() torch.nn.utils.clip_grad_norm_(head.parameters(), 1.0) optim.step() lam_mean_val = mean_lyap.detach().mean().item() rec = {"step": step, "sup_loss": float(sup_loss.item()), "rf_loss": float(rf_loss.item()), "total_loss": float(total_loss.item()), "mean_lyap": lam_mean_val, "lam_star": float(lam_star.mean().item() if lam_star.dim() else lam_star.item()), "excess_frac_nonzero": float((excess > 0).float().mean().item())} log["steps"].append(rec) if step % 5 == 0 or step == args.n_steps - 1: print(f" [{step:>4}/{args.n_steps}] dt={time.time()-t0:.1f}s " f"sup={rec['sup_loss']:.4f} rf={rec['rf_loss']:.4f} " f"lyap={lam_mean_val:+.4f} λ*={rec['lam_star']:+.4f} " f"nz={rec['excess_frac_nonzero']:.2f}", flush=True) if (step + 1) % args.eval_every == 0: acc, tacc = evaluate(head, base, Path(cfg["data_path"]), args.eval_n, args.eval_batch_size, device) print(f" >> EVAL step {step+1}: exact_acc={acc:.4f} (Δ from init: {acc-acc0:+.4f})") log["steps"][-1]["eval_acc"] = acc log["steps"][-1]["eval_tok_acc"] = tacc head.train() print(f"\n=== Final eval ===") acc_f, tacc_f = evaluate(head, base, Path(cfg["data_path"]), args.eval_n, args.eval_batch_size, device) print(f" initial: {acc0:.4f} → final: {acc_f:.4f} (Δ {acc_f-acc0:+.4f})") log["final_acc"] = acc_f; log["final_tok_acc"] = tacc_f Path(args.out).write_text(json.dumps(log, indent=2)) print(f"log saved → {args.out}") if __name__ == "__main__": main()