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
|
"""(a) lite: For each test sample, save the final tangent basis Q (top-k modes after
running through the full inference). Compute position/hidden activity profiles per
mode and compare success vs failure groups.
"""
from __future__ import annotations
import sys, os, yaml, json, time, argparse
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, ckpt_name, device):
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 = {}
for k, v in sd.items():
nk = k
for p in ("_orig_mod.", "model."):
if nk.startswith(p): nk = nk[len(p):]
stripped[nk] = v
model.load_state_dict(stripped, strict=False)
model.to(device).eval()
return model, cfg, train_meta
def jvp_one(f, x, v):
return torch.autograd.functional.jvp(f, x, v=v, create_graph=False, strict=False)
def run_save_final_Q(model, batch, k_lyap, device, seed):
"""Run inference with QR-iteration on top-k tangents; return final Q (B, seq, hidden, k)
after all ACT steps. Also return exact_correct, predicted_logits.
"""
inner = model.inner
cfg = inner.config
B = batch["inputs"].shape[0]
seq_full = cfg.seq_len + inner.puzzle_emb_len
hidden = cfg.hidden_size
state_dim = 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))
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)
with torch.enable_grad():
for _act in range(cfg.halt_max_steps):
zH = z_H.detach(); zL = z_L.detach()
for _h in range(cfg.H_cycles):
for _l in range(cfg.L_cycles):
out = []
fx_last = None
f = lambda x: inner.L_level(x, zH + input_embeddings, **seq_info)
for i in range(k_lyap):
v_i = Q[..., i].view_as(zL)
fx, Dv = jvp_one(f, zL, v_i)
out.append(Dv.reshape(B, state_dim).to(torch.float32))
fx_last = fx
Q = torch.stack(out, dim=-1)
zL = fx_last
Q, R = torch.linalg.qr(Q)
out = []
f = lambda x: inner.H_level(x, zL, **seq_info)
for i in range(k_lyap):
v_i = Q[..., i].view_as(zH)
fx, Dv = jvp_one(f, zH, v_i)
out.append(Dv.reshape(B, state_dim).to(torch.float32))
fx_last = fx
Q = torch.stack(out, dim=-1)
zH = fx_last
Q, R = torch.linalg.qr(Q)
z_H, z_L = zH, zL
with torch.no_grad():
output = inner.lm_head(z_H)[:, inner.puzzle_emb_len:].float()
preds = output.argmax(dim=-1)
labels = batch["labels"].to(device)
mask = labels > 0
exact = ((preds == labels) | ~mask).all(dim=-1).cpu().float().numpy()
Q_final = Q.reshape(B, seq_full, hidden, k_lyap).cpu().float().numpy()
return Q_final, exact
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=256)
ap.add_argument("--batch-size", type=int, default=32)
ap.add_argument("--k-lyap", type=int, default=4)
ap.add_argument("--seed", type=int, default=0)
ap.add_argument("--out", default="tangent_modes.npz")
args = ap.parse_args()
device = "cuda"
model, cfg, train_meta = load_model(Path(args.ckpt_root), args.ckpt_name, device)
rng = np.random.default_rng(args.seed)
data_path = Path(cfg["data_path"])
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=args.n_samples, replace=False)
Q_all = []; exact_all = []
t0 = time.time()
for s in range(0, args.n_samples, args.batch_size):
e = min(s + args.batch_size, args.n_samples)
bidx = idx[s:e]
batch = {
"inputs": torch.from_numpy(inputs[bidx].astype(np.int32)),
"labels": torch.from_numpy(labels[bidx].astype(np.int32)),
"puzzle_identifiers": torch.from_numpy(pid[bidx].astype(np.int32)),
}
Q_final, exact = run_save_final_Q(model, batch, args.k_lyap, device, seed=args.seed + s)
Q_all.append(Q_final); exact_all.append(exact)
print(f" [{e}/{args.n_samples}] dt={time.time()-t0:.1f}s exact={exact.mean():.3f}", flush=True)
Q_all = np.concatenate(Q_all, axis=0) # (N, seq, hidden, k)
exact_all = np.concatenate(exact_all, axis=0) # (N,)
print(f"saved shape Q={Q_all.shape}, exact={exact_all.shape}, acc={exact_all.mean():.3f}")
np.savez_compressed(args.out, Q_final=Q_all, exact_correct=exact_all, sample_idx=idx)
print(f"saved → {args.out}")
if __name__ == "__main__":
main()
|
