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
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
|
"""HRM diagnostic: separate λ_L (固定 z_H 下 L 子系统) and λ_H (z_L 收敛后 H 子系统),
in addition to the joint λ from diagnose_hrm_joint.py.
Three orthonormal bases evolved in parallel:
- Q_joint: (B, 2D, k) — joint (v_H, v_L). Block-matrix update per L/H step.
- Q_L: (B, D, k) — only updated during L steps via J_L. Unchanged during H steps.
- Q_H: (B, D, k) — only updated during H steps via J_H. Unchanged during L steps.
Note: Q_L's evolution uses J_L evaluated at the current trajectory's z_L+z_H+ie,
which means we measure the L sub-system Lyapunov "along the actual z_H trajectory"
(z_H changes between H-cycles). Similarly for Q_H.
"""
from __future__ import annotations
import sys, os, yaml, math, argparse, json, time
from pathlib import Path
import numpy as np
import torch
HRM_DIR = Path("/home/yurenh2/rrm/hrm")
sys.path.insert(0, str(HRM_DIR))
from models.hrm.hrm_act_v1 import HierarchicalReasoningModel_ACTV1
def load_model(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)
model = HierarchicalReasoningModel_ACTV1(arch_cfg)
sd = torch.load(ckpt_root / ckpt_name, map_location="cpu", weights_only=True)
stripped = {k.replace("_orig_mod.", "").replace("model.", ""): v for k, v in sd.items()}
model.load_state_dict(stripped, strict=False)
model.to(device).eval()
return model, cfg, train_meta
def load_test_samples(data_path, n_total, shard_id, num_shards, seed):
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")
all_idx = rng.choice(len(inputs), size=n_total, replace=False)
shard_size = (n_total + num_shards - 1) // num_shards
s, e = shard_id * shard_size, min((shard_id + 1) * shard_size, n_total)
idx = all_idx[s:e]
return {
"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)),
"idx": idx,
}
def jvp(f, x, v):
return torch.autograd.functional.jvp(f, x, v=v, create_graph=False, strict=False)
def run_diagnose_batch(model, batch, device, k_lyap, t_ons, seed):
inner = model.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"].to(device),
batch["puzzle_identifiers"].to(device))
# Three independent orthonormal bases
g = torch.Generator(device=device).manual_seed(seed)
Q_joint = torch.linalg.qr(torch.randn(B, 2*D, k_lyap, device=device, dtype=torch.float32, generator=g))[0]
Q_L = torch.linalg.qr(torch.randn(B, D, k_lyap, device=device, dtype=torch.float32, generator=g))[0]
Q_H = torch.linalg.qr(torch.randn(B, D, k_lyap, device=device, dtype=torch.float32, generator=g))[0]
log_R_joint = torch.zeros(B, k_lyap, device=device, dtype=torch.float32)
log_R_L = torch.zeros(B, k_lyap, device=device, dtype=torch.float32)
log_R_H = torch.zeros(B, k_lyap, device=device, dtype=torch.float32)
n_joint_steps = 0; n_L_steps = 0; n_H_steps = 0
step_counter_joint = 0; step_counter_L = 0; step_counter_H = 0
for act_step in range(cfg.halt_max_steps):
with torch.enable_grad():
zH, zL = z_H.detach(), z_L.detach()
for _h in range(cfg.H_cycles):
for _l in range(cfg.L_cycles):
# ============ JOINT update (L step) ============
v_H_j = Q_joint[:, :D, :]
v_L_j = Q_joint[:, D:, :]
v_comb = v_H_j + v_L_j
new_v_L_j_cols = []
f_L = lambda z: inner.L_level(z, zH + input_embeddings, **seq_info)
for i in range(k_lyap):
v_i = v_comb[:, :, i].reshape(B, seq_full, hidden).to(inner.forward_dtype)
zL_new, Dv = jvp(f_L, zL, v_i)
new_v_L_j_cols.append(Dv.reshape(B, D).to(torch.float32))
new_v_L_j = torch.stack(new_v_L_j_cols, dim=-1)
Q_joint = torch.cat([v_H_j, new_v_L_j], dim=1)
# ============ L-only update ============
new_v_L_only_cols = []
for i in range(k_lyap):
v_i = Q_L[:, :, i].reshape(B, seq_full, hidden).to(inner.forward_dtype)
_, Dv = jvp(f_L, zL, v_i)
new_v_L_only_cols.append(Dv.reshape(B, D).to(torch.float32))
Q_L = torch.stack(new_v_L_only_cols, dim=-1)
# Q_H untouched during L step (since H_level wasn't applied)
zL = zL_new
step_counter_joint += 1; step_counter_L += 1
if step_counter_joint % t_ons == 0:
Q_joint, Rj = torch.linalg.qr(Q_joint)
log_R_joint = log_R_joint + Rj.diagonal(dim1=-2, dim2=-1).abs().clamp_min(1e-30).log()
n_joint_steps += 1
if step_counter_L % t_ons == 0:
Q_L, Rl = torch.linalg.qr(Q_L)
log_R_L = log_R_L + Rl.diagonal(dim1=-2, dim2=-1).abs().clamp_min(1e-30).log()
n_L_steps += 1
# ============ JOINT update (H step) ============
v_H_j = Q_joint[:, :D, :]
v_L_j = Q_joint[:, D:, :]
v_comb = v_H_j + v_L_j
new_v_H_j_cols = []
f_H = lambda z: inner.H_level(z, zL, **seq_info)
for i in range(k_lyap):
v_i = v_comb[:, :, i].reshape(B, seq_full, hidden).to(inner.forward_dtype)
zH_new, Dv = jvp(f_H, zH, v_i)
new_v_H_j_cols.append(Dv.reshape(B, D).to(torch.float32))
new_v_H_j = torch.stack(new_v_H_j_cols, dim=-1)
Q_joint = torch.cat([new_v_H_j, v_L_j], dim=1)
# ============ H-only update ============
new_v_H_only_cols = []
for i in range(k_lyap):
v_i = Q_H[:, :, i].reshape(B, seq_full, hidden).to(inner.forward_dtype)
_, Dv = jvp(f_H, zH, v_i)
new_v_H_only_cols.append(Dv.reshape(B, D).to(torch.float32))
Q_H = torch.stack(new_v_H_only_cols, dim=-1)
# Q_L untouched during H step
zH = zH_new
step_counter_joint += 1; step_counter_H += 1
if step_counter_joint % t_ons == 0:
Q_joint, Rj = torch.linalg.qr(Q_joint)
log_R_joint = log_R_joint + Rj.diagonal(dim1=-2, dim2=-1).abs().clamp_min(1e-30).log()
n_joint_steps += 1
if step_counter_H % t_ons == 0:
Q_H, Rh = torch.linalg.qr(Q_H)
log_R_H = log_R_H + Rh.diagonal(dim1=-2, dim2=-1).abs().clamp_min(1e-30).log()
n_H_steps += 1
z_H, z_L = zH, zL
with torch.no_grad():
output = inner.lm_head(z_H)[:, inner.puzzle_emb_len:].float()
final_logits = output
lyap_joint = (log_R_joint / max(n_joint_steps, 1)).cpu().numpy()
lyap_L = (log_R_L / max(n_L_steps, 1)).cpu().numpy()
lyap_H = (log_R_H / max(n_H_steps, 1)).cpu().numpy()
with torch.no_grad():
preds = final_logits.argmax(dim=-1)
labels = batch["labels"].to(device)
mask = labels > 0
exact = ((preds == labels) | ~mask).all(dim=-1).cpu().float()
token_acc = ((preds == labels) & mask).sum(-1).float() / mask.sum(-1).float().clamp_min(1)
token_acc = token_acc.cpu()
return {
"lyap_joint": lyap_joint,
"lyap_L": lyap_L,
"lyap_H": lyap_H,
"exact_correct": exact.numpy(),
"token_acc": token_acc.numpy(),
}
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--ckpt-root", required=True)
ap.add_argument("--ckpt-name", default="step_26040")
ap.add_argument("--n-samples", type=int, default=1024)
ap.add_argument("--shard-id", type=int, default=0)
ap.add_argument("--num-shards", type=int, default=1)
ap.add_argument("--batch-size", type=int, default=32)
ap.add_argument("--k-lyap", type=int, default=8)
ap.add_argument("--t-ons", type=int, default=1)
ap.add_argument("--seed", type=int, default=0)
ap.add_argument("--out", default="diag_separate.npz")
args = ap.parse_args()
device = "cuda"
model, cfg, train_meta = load_model(Path(args.ckpt_root), args.ckpt_name, device)
test = load_test_samples(Path(cfg["data_path"]), args.n_samples,
args.shard_id, args.num_shards, args.seed)
n = len(test["inputs"])
print(f"shard {args.shard_id}/{args.num_shards}: {n} samples")
print(f"H_cycles={model.inner.config.H_cycles} L_cycles={model.inner.config.L_cycles} halt={model.inner.config.halt_max_steps}")
res = {k: [] for k in ["lyap_joint","lyap_L","lyap_H","exact_correct","token_acc","idx"]}
t0 = time.time()
for s in range(0, n, args.batch_size):
e = min(s + args.batch_size, n)
batch = {k: test[k][s:e].to(device) for k in ["inputs","labels","puzzle_identifiers"]}
out = run_diagnose_batch(model, batch, device, args.k_lyap, args.t_ons, args.seed + s)
for k, v in out.items():
res[k].append(v)
res["idx"].append(test["idx"][s:e])
print(f" [{e}/{n}] dt={time.time()-t0:.1f}s exact={out['exact_correct'].mean():.3f} "
f"λj1={out['lyap_joint'][:,0].mean():+.3f} "
f"λL1={out['lyap_L'][:,0].mean():+.3f} "
f"λH1={out['lyap_H'][:,0].mean():+.3f}", flush=True)
saved = {}
for k, v in res.items():
if not v: continue
try: saved[k] = np.concatenate(v, 0)
except ValueError: saved[k] = np.stack(v, 0)
np.savez_compressed(args.out, **saved)
succ = saved["exact_correct"] > 0.5
print(f"\nN={len(succ)} acc={succ.mean():.4f}")
for name in ["lyap_joint", "lyap_L", "lyap_H"]:
ls = saved[name]
print(f"\n{name}:")
print(f" i mean_succ mean_fail Δ")
for i in range(ls.shape[1]):
ms, mf = ls[succ,i].mean(), ls[~succ,i].mean()
print(f" {i+1} {ms:+8.4f} {mf:+8.4f} {mf-ms:+8.4f}")
print(f"\nsaved → {args.out}")
if __name__ == "__main__":
main()
|