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
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
|
"""Empirical Lipschitz sanity check: perturb init state by small noise,
measure how OUTPUT and final z_H change. Independent of our JVP code.
If TRM succ samples truly have λ > 0, perturbations should diverge through dynamics.
If they're actually stable in output subspace, perturbations decay or stay bounded.
"""
import sys, yaml, json, math
from pathlib import Path
import numpy as np
import torch
HRM_DIR = Path("/home/yurenh2/rrm/hrm")
TRM_DIR = Path("/home/yurenh2/rrm/trm")
CKPT_TRM_ROOT = "/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_singleGPU"
CKPT_TRM_NAME = "step_104164"
CKPT_HRM_ROOT = "/home/yurenh2/rrm/hrm/checkpoints/Sudoku-extreme-1k-aug-1000 ACT-torch/HierarchicalReasoningModel_ACTV1 righteous-python"
CKPT_HRM_NAME = "step_26040"
DEVICE = "cuda"
def load_model(repo_dir, ckpt_root, ckpt_name, model_cls_path):
# Clear cached modules from other repo to avoid conflicts (HRM/TRM both have models.*)
for mod in list(sys.modules.keys()):
if mod.startswith("models"):
del sys.modules[mod]
sys.path[:] = [p for p in sys.path if not (p.endswith("/hrm") or p.endswith("/trm"))]
sys.path.insert(0, str(repo_dir))
import importlib
mod_path, cls_name = model_cls_path.split("@")
cls = getattr(importlib.import_module(mod_path), cls_name)
cfg = yaml.safe_load((Path(ckpt_root) / "all_config.yaml").read_text())
arch_cfg = dict(cfg["arch"])
data_path = Path(cfg.get("data_path") or cfg["data_paths"][0])
train_meta = json.loads((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)
model = cls(arch_cfg)
sd = torch.load(Path(ckpt_root) / ckpt_name, map_location="cpu", weights_only=True)
sd = {k.replace("_orig_mod.", "").replace("model.", ""): v for k, v in sd.items()}
missing, unexpected = model.load_state_dict(sd, strict=False)
print(f" [load] missing={len(missing)} unexpected={len(unexpected)}")
model.to(DEVICE).eval()
return model, cfg, train_meta, data_path
def load_test_samples(data_path, n_total, seed=0):
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 = rng.choice(len(inputs), size=n_total, replace=False)
return {
"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),
}
@torch.no_grad()
def measure_pert_stability(model, batch, eps=1e-2, n_act_steps=8):
"""For each sample, run UNPERTURBED + PERTURBED full forward.
Track:
- δz_H (final): norm of z_H change at end of all ACT steps
- δz_L (final): same for z_L
- argmax flip: did the prediction change?
Returns per-sample stats.
The "growth rate" inferred from δz_final / δz_init can be compared to JVP λ.
If λ_JVP > 0, δz_final >> δz_init (expansion). If λ_JVP < 0, δz_final < δz_init.
"""
inner = model.inner
cfg = inner.config
B = batch["inputs"].shape[0]
seq_full = cfg.seq_len + inner.puzzle_emb_len
hidden = cfg.hidden_size
dt = inner.forward_dtype
# Initial z_H, z_L (identical for both runs initially)
z_H_0 = inner.H_init.unsqueeze(0).expand(B, seq_full, hidden).clone().to(dt)
z_L_0 = inner.L_init.unsqueeze(0).expand(B, seq_full, hidden).clone().to(dt)
# Perturbation
g = torch.Generator(device=DEVICE).manual_seed(42)
delta_H = torch.randn(B, seq_full, hidden, generator=g, dtype=torch.float32, device=DEVICE).to(dt) * eps
delta_L = torch.randn(B, seq_full, hidden, generator=g, dtype=torch.float32, device=DEVICE).to(dt) * eps
input_emb = inner._input_embeddings(batch["inputs"], batch["puzzle_identifiers"])
seq_info = dict(cos_sin=inner.rotary_emb() if hasattr(inner, "rotary_emb") else None)
init_delta_norm = (delta_H.float().flatten(1).norm(dim=1) +
delta_L.float().flatten(1).norm(dim=1)) # (B,) sum of init pert norms
# Run unperturbed and perturbed in parallel
z_H_a, z_L_a = z_H_0.clone(), z_L_0.clone()
z_H_b, z_L_b = z_H_0 + delta_H, z_L_0 + delta_L
has_H_level = hasattr(inner, "H_level") # HRM has separate, TRM uses L_level for both
n_total = 0
for _act in range(n_act_steps):
for _h in range(cfg.H_cycles):
for _l in range(cfg.L_cycles):
z_L_a = inner.L_level(z_L_a, z_H_a + input_emb, **seq_info)
z_L_b = inner.L_level(z_L_b, z_H_b + input_emb, **seq_info)
n_total += 1
# H step: use H_level (HRM) or L_level (TRM)
h_mod = inner.H_level if has_H_level else inner.L_level
z_H_a = h_mod(z_H_a, z_L_a, **seq_info)
z_H_b = h_mod(z_H_b, z_L_b, **seq_info)
n_total += 1
final_delta_norm = ((z_H_b - z_H_a).float().flatten(1).norm(dim=1) +
(z_L_b - z_L_a).float().flatten(1).norm(dim=1))
# Per-sample growth rate per micro-step
# δ_final ≈ δ_init * exp(λ * n_total) → λ ≈ log(δ_final/δ_init) / n_total
ratio = final_delta_norm / init_delta_norm.clamp_min(1e-12)
lam_emp = ratio.log() / n_total
# Read out predictions for both runs
out_a = inner.lm_head(z_H_a)[:, inner.puzzle_emb_len:].float()
out_b = inner.lm_head(z_H_b)[:, inner.puzzle_emb_len:].float()
pred_a = out_a.argmax(dim=-1)
pred_b = out_b.argmax(dim=-1)
labels = batch["labels"]
mask = labels > 0
exact_a = ((pred_a == labels) | ~mask).all(dim=-1)
exact_b = ((pred_b == labels) | ~mask).all(dim=-1)
pred_flip = (pred_a != pred_b).any(dim=-1) # any token changed
return {
"init_norm": init_delta_norm.cpu(),
"final_norm": final_delta_norm.cpu(),
"ratio": ratio.cpu(),
"lam_emp": lam_emp.cpu(),
"succ_a": exact_a.cpu(),
"succ_b": exact_b.cpu(),
"pred_flip": pred_flip.cpu(),
}
def main():
for name, repo, ckpt_root, ckpt_name, mod_path in [
("HRM step_26040", HRM_DIR, CKPT_HRM_ROOT, CKPT_HRM_NAME,
"models.hrm.hrm_act_v1@HierarchicalReasoningModel_ACTV1"),
("TRM step_104164", TRM_DIR, CKPT_TRM_ROOT, CKPT_TRM_NAME,
"models.recursive_reasoning.trm@TinyRecursiveReasoningModel_ACTV1"),
]:
print(f"\n=== {name} ===")
model, cfg, train_meta, data_path = load_model(repo, ckpt_root, ckpt_name, mod_path)
batch = load_test_samples(data_path, n_total=64, seed=0)
# Limit batch size to model's training batch (puzzle_emb buffer)
# Re-batch
B = 16
results = {"lam_emp": [], "succ_a": [], "ratio": []}
for s in range(0, 64, B):
e = min(s + B, 64)
mb = {k: v[s:e] for k, v in batch.items()}
# Rebuild model puzzle_emb buffer if needed — easier: ensure model's batch_size matches
r = measure_pert_stability(model, mb, eps=1e-2, n_act_steps=8)
results["lam_emp"].append(r["lam_emp"])
results["succ_a"].append(r["succ_a"])
results["ratio"].append(r["ratio"])
lam = torch.cat(results["lam_emp"]).numpy()
succ = torch.cat(results["succ_a"]).numpy()
ratio = torch.cat(results["ratio"]).numpy()
print(f" N=64 acc={succ.mean():.3f}")
print(f" finite-diff λ_emp (per micro-step):")
print(f" all mean={lam.mean():+.4f} med={np.median(lam):+.4f} range=[{lam.min():+.4f}, {lam.max():+.4f}]")
if succ.sum() > 0:
print(f" succ mean={lam[succ.astype(bool)].mean():+.4f} med={np.median(lam[succ.astype(bool)]):+.4f}")
if (~succ).sum() > 0:
print(f" fail mean={lam[~succ.astype(bool)].mean():+.4f} med={np.median(lam[~succ.astype(bool)]):+.4f}")
print(f" final/init perturbation ratio:")
print(f" all mean={ratio.mean():.3f} med={np.median(ratio):.3f} range=[{ratio.min():.3e}, {ratio.max():.3e}]")
# cleanup
del model
torch.cuda.empty_cache()
if __name__ == "__main__":
main()
|
