"""Step 4: From-scratch HRM training with CF regularizer (no checkpoint load). Tests whether forcing λ_joint_1 → λ* from step 0 affects learning trajectory. Hypothesis test: - Baseline (α=0): λ naturally drifts toward HRM's attractor (~-0.15) - CF λ*=0: forces λ to stay near 0 (edge of chaos) - CF λ*=-0.15: enforces natural attractor from start (should be neutral) For each condition we track λ trajectory + acc + halt distribution at fixed steps. """ 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 from models.losses import ACTLossHead from models.sparse_embedding import CastedSparseEmbeddingSignSGD_Distributed from adam_atan2 import AdamATan2 def build_model_from_scratch(data_path: Path, batch_size: int, device: str, hidden_size: int = 512, num_heads: int = 8): """Build HRM with the official Sudoku-1k arch config but at our batch size + arbitrary hidden.""" train_meta = json.loads((data_path / "train" / "dataset.json").read_text()) arch_cfg = dict( H_cycles=2, H_layers=4, L_cycles=2, L_layers=4, expansion=4, halt_exploration_prob=0.1, halt_max_steps=16, hidden_size=hidden_size, num_heads=num_heads, pos_encodings="rope", puzzle_emb_ndim=hidden_size, loss=dict(loss_type="stablemax_cross_entropy", name="losses@ACTLossHead"), batch_size=batch_size, vocab_size=train_meta["vocab_size"], seq_len=train_meta["seq_len"], num_puzzle_identifiers=train_meta["num_puzzle_identifiers"], causal=False, ) with torch.device(device): base = HierarchicalReasoningModel_ACTV1(arch_cfg) head = ACTLossHead(base, loss_type=arch_cfg["loss"]["loss_type"]) return head, base, train_meta def jvp_train(f, x, v): return torch.autograd.functional.jvp(f, x, v=v, create_graph=True, strict=False) def compute_joint_lyap_spec(base, batch, k_lyap, lyap_act_steps, device, seed, with_grad=True): """Returns FULL top-k Lyapunov spectrum (B, k), differentiable wrt theta.""" inner = base.inner cfg = inner.config B = batch["inputs"].shape[0] seq_full = cfg.seq_len + inner.puzzle_emb_len hidden = cfg.hidden_size D = seq_full * hidden 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, 2*D, 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 jvp_fn = (lambda f, x, v: torch.autograd.functional.jvp(f, x, v=v, create_graph=with_grad, strict=False)) 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): v_H_j = Q[:, :D, :]; v_L_j = Q[:, D:, :] v_comb = v_H_j + v_L_j new_v_L_cols = [] f_L = lambda z: inner.L_level(z, z_H + input_embeddings, **seq_info) for i in range(k_lyap): v_i = v_comb[:, :, i].reshape(B, seq_full, hidden).to(inner.forward_dtype) z_L_new, Dv = jvp_fn(f_L, z_L, v_i) new_v_L_cols.append(Dv.reshape(B, D).to(torch.float32)) new_v_L = torch.stack(new_v_L_cols, dim=-1) Q = torch.cat([v_H_j, new_v_L], dim=1) 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 v_H_j = Q[:, :D, :]; v_L_j = Q[:, D:, :] v_comb = v_H_j + v_L_j new_v_H_cols = [] f_H = lambda z: inner.H_level(z, z_L, **seq_info) for i in range(k_lyap): v_i = v_comb[:, :, i].reshape(B, seq_full, hidden).to(inner.forward_dtype) z_H_new, Dv = jvp_fn(f_H, z_H, v_i) new_v_H_cols.append(Dv.reshape(B, D).to(torch.float32)) new_v_H = torch.stack(new_v_H_cols, dim=-1) Q = torch.cat([new_v_H, v_L_j], dim=1) 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) return lyap_spec # (B, k_lyap) def load_train_batches(data_path: Path, batch_size: int, n_iters: int, seed: int = 0): 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, n_samples, batch_size, device, seed=42): 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) def warmup_constant_lr(step, base_lr, warmup): if step < warmup: return base_lr * step / max(1, warmup) return base_lr def main(): ap = argparse.ArgumentParser() ap.add_argument("--data-path", default="/home/yurenh2/rrm/data/sudoku-extreme-1k-aug-1000") ap.add_argument("--n-steps", type=int, default=2500) ap.add_argument("--batch-size", type=int, default=8) ap.add_argument("--lr", type=float, default=1e-4) ap.add_argument("--puzzle-emb-lr", type=float, default=1e-4) ap.add_argument("--warmup-steps", type=int, default=200) ap.add_argument("--weight-decay", type=float, default=1.0) ap.add_argument("--hidden-size", type=int, default=512, help="HRM hidden dim (default official 512)") ap.add_argument("--num-heads", type=int, default=8, help="must divide hidden_size; default 8") ap.add_argument("--alpha-rf", type=float, default=0.0, help="0 = baseline; >0 = CF/Engelken") ap.add_argument("--rf-mode", choices=["fixed", "volume_cf", "engelken_l2"], default="fixed", help="fixed: hinge max(0, λ_1-λ*)² on top-1; " "volume_cf: hinge max(0, mean_i λ_i-λ*)² over top-k; " "engelken_l2: (1/k) Σ λ_i² across full top-k spectrum") ap.add_argument("--lambda-star", type=float, default=0.0, help="used in fixed and volume_cf modes") ap.add_argument("--k-lyap", type=int, default=2) ap.add_argument("--lyap-act-steps", type=int, default=4) ap.add_argument("--seed", type=int, default=42) ap.add_argument("--eval-every", type=int, default=250) ap.add_argument("--eval-n", type=int, default=512) ap.add_argument("--eval-batch-size", type=int, default=32) ap.add_argument("--out", required=True) ap.add_argument("--save-ckpt", default="", help="path to save final model state_dict (empty = skip)") args = ap.parse_args() device = "cuda" torch.manual_seed(args.seed); np.random.seed(args.seed) data_path = Path(args.data_path) head, base, train_meta = build_model_from_scratch(data_path, args.batch_size, device, hidden_size=args.hidden_size, num_heads=args.num_heads) print(f"Built HRM from scratch | params={sum(p.numel() for p in head.parameters()):,} | " f"vocab={train_meta['vocab_size']} seq={train_meta['seq_len']} " f"num_pids={train_meta['num_puzzle_identifiers']}") # Two optimizers: SignSGD for puzzle_emb (sparse), AdamATan2 for rest puzzle_emb_opt = CastedSparseEmbeddingSignSGD_Distributed( base.inner.puzzle_emb.buffers(), lr=0, weight_decay=args.weight_decay, world_size=1, ) main_opt = AdamATan2(head.parameters(), lr=0, betas=(0.9, 0.95), weight_decay=args.weight_decay) # Baseline eval (random init) acc0, tacc0 = evaluate(head, base, data_path, args.eval_n, args.eval_batch_size, device) print(f"=== step 0 (random init): exact_acc = {acc0:.4f} token_acc = {tacc0:.4f} ===") log = {"args": vars(args), "initial_acc": acc0, "initial_tok_acc": tacc0, "steps": [], "evals": []} log["evals"].append({"step": 0, "acc": acc0, "tok_acc": tacc0}) t0 = time.time() train_iter = load_train_batches(data_path, args.batch_size, args.n_steps, seed=args.seed) for step, batch in enumerate(train_iter): batch = {k: v.to(device) for k, v in batch.items()} # Update LR cur_lr = warmup_constant_lr(step, args.lr, args.warmup_steps) cur_pe_lr = warmup_constant_lr(step, args.puzzle_emb_lr, args.warmup_steps) for pg in main_opt.param_groups: pg["lr"] = cur_lr for pg in puzzle_emb_opt.param_groups: pg["lr"] = cur_pe_lr head.train() # ACT loss 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 sup_loss = sup_loss_sum / max(n_loss, 1) / args.batch_size # CF / Engelken loss (skip if alpha=0; still measure λ for logging in baseline) if args.alpha_rf > 0: lyap_spec = compute_joint_lyap_spec(base, batch, args.k_lyap, args.lyap_act_steps, device, args.seed + step, with_grad=True) lyap1 = lyap_spec[:, 0] if args.rf_mode == "engelken_l2": # Engelken: L = (1/k) Σ_i λ_i² over batch → push all spectrum toward 0 rf_loss = (lyap_spec ** 2).mean() excess = lyap1 # for logging (no hinge here) elif args.rf_mode == "volume_cf": # One-sided cap on local phase-space volume expansion. # Allows λ_1 > 0 when compensated by contraction in other measured modes. lyap_volume = lyap_spec.mean(dim=1) excess = (lyap_volume - args.lambda_star).clamp_min(0.0) rf_loss = (excess ** 2).mean() else: # fixed (hinge on top-1) excess = (lyap1 - args.lambda_star).clamp_min(0.0) rf_loss = (excess ** 2).mean() else: with torch.no_grad(): lyap_spec = compute_joint_lyap_spec(base, batch, args.k_lyap, args.lyap_act_steps, device, args.seed + step, with_grad=False) lyap1 = lyap_spec[:, 0] rf_loss = torch.zeros((), device=device) excess = (lyap1 - args.lambda_star).clamp_min(0.0) total_loss = sup_loss + args.alpha_rf * rf_loss puzzle_emb_opt.zero_grad(set_to_none=True) main_opt.zero_grad(set_to_none=True) total_loss.backward() torch.nn.utils.clip_grad_norm_([p for p in head.parameters() if p.requires_grad], 1.0) main_opt.step() puzzle_emb_opt.step() with torch.no_grad(): lyap_mean_per_i = lyap_spec.detach().mean(dim=0).cpu().tolist() rec = { "step": step, "lr": cur_lr, "sup_loss": float(sup_loss.item()), "rf_loss": float(rf_loss.item()), "total_loss": float(total_loss.item()), "lyap1_mean": float(lyap1.detach().mean().item()), "lyap1_max": float(lyap1.detach().max().item()), "lyap1_min": float(lyap1.detach().min().item()), "lyap_volume_mean": float(lyap_spec.detach().mean(dim=1).mean().item()), "lyap_volume_max": float(lyap_spec.detach().mean(dim=1).max().item()), "lyap_spec_mean": lyap_mean_per_i, "frac_above_star": float((excess > 0).float().mean().item()), } log["steps"].append(rec) if step % 25 == 0 or step == args.n_steps - 1: print(f" [{step:>4}/{args.n_steps}] dt={time.time()-t0:.0f}s lr={cur_lr:.1e} " f"sup={rec['sup_loss']:.4f} rf={rec['rf_loss']:.4f} " f"λ_mean={rec['lyap1_mean']:+.4f} [{rec['lyap1_min']:+.3f},{rec['lyap1_max']:+.3f}] " f"frac>λ*={rec['frac_above_star']:.2f}", flush=True) if (step + 1) % args.eval_every == 0 or step == args.n_steps - 1: acc, tacc = evaluate(head, base, data_path, args.eval_n, args.eval_batch_size, device) print(f" >> EVAL @ {step+1}: exact_acc={acc:.4f} tok_acc={tacc:.4f} " f"(Δ from init: {acc-acc0:+.4f})", flush=True) log["evals"].append({"step": step + 1, "acc": acc, "tok_acc": tacc}) log["final_acc"] = log["evals"][-1]["acc"] log["final_tok_acc"] = log["evals"][-1]["tok_acc"] Path(args.out).write_text(json.dumps(log, indent=2)) print(f"\n=== DONE === init {acc0:.4f} → final {log['final_acc']:.4f} log → {args.out}") if args.save_ckpt: save_path = Path(args.save_ckpt) save_path.parent.mkdir(parents=True, exist_ok=True) torch.save({ "state_dict": head.state_dict(), "args": vars(args), "n_steps_trained": args.n_steps, "final_acc": log["final_acc"], "final_tok_acc": log["final_tok_acc"], "arch_cfg_signature": { "vocab_size": train_meta["vocab_size"], "seq_len": train_meta["seq_len"], "num_puzzle_identifiers": train_meta["num_puzzle_identifiers"], "batch_size": args.batch_size, }, }, save_path) print(f"checkpoint saved → {save_path}") if __name__ == "__main__": main()